CN105426911A - Dirichlet process mixture model based TAC clustering method - Google Patents

Abstract

The invention discloses a Dirichlet process mixture model based TAC clustering method. The method comprises: initializing a Dirichlet process mixture model; iteratively calculating a conditional probability; and performing sampling until an iterative stop condition is met. According to the method, the TAC clustering is performed by using the Dirichlet process mixture model, so that the problem in TAC clustering performed under the condition of an unknown class number is effectively solved; and compared with other clustering algorithms, the method has the advantage that the complexity of the Dirichlet process mixture model can be increased with an increase in volume of obtained data.

Description

A kind of TAC clustering method based on Di Li Cray process mixture model

Technical field

The invention belongs to clustering technique field, be specifically related to a kind of TAC clustering method based on Di Li Cray process mixture model.

Background technology

Positron emission tomography imaging (positronemissiontomography, PET) is a kind of medicine imaging technique.Dynamic pet imaging is gathered by continuous data, obtain multiframe physiological status space distribution, it can by rebuilding the distribution on the Time and place of the bio-matrix of radiopharmaceutical agent mark in biological tissue, provide about different biologies or physiology course quantification and the information of non-intrusion type.In practice, dynamic PET images is often divided into different area-of-interests (regionofinterest, ROI), then from each region, extracts time activity curve (timeactivitycurve, TAC).TAC analyzedly further can estimate physiologic parameters, and as volume of blood flow, metabolism and acceptor density, this depends on the characteristic of tracer agent.

Only have several ROI based on dynamic PET images and each ROI is that uniform this is true, we wish to carry out cluster to TAC, namely split dynamic PET images.More importantly, we wish to carry out cluster to TAC by feature based, make noise robust more.Di Li Cray process (DirichletProcess, DP) is the representational stochastic process of most in nonparametric bayes method.Different from other clustering methods, Di Li Cray process mixture model can carry out cluster when not knowing data class number to data, and along with the getable data grows of model many, it adaptive characteristic according to data self can carry out cluster to data and provides the information of data class number.

In the field of machine learning, probability model is used to the distribution of modeling based on observation data.Traditional parameter model uses fixing and a limited number of parameter, time this makes unmatched between the complexity (often weighing by the number of parameter) and available data volume of model, traditional parameter model is easy to suffer over-fitting or poor fitting.Therefore, the selection of model, has the selection of the model of correct complexity in other words, is very important problem in parameter model.Then, no matter we are that the selection of model is all very difficult using cross validation or marginal probability as the basis selected.Bayes's nonparametric technique is an option of parameter model and selection, it has the model of unlimited complexity by use one, the situation of poor fitting can alleviate, and meanwhile, uses bayes method calculating and the approximate complete posteriority based on parameter to alleviate the situation of over-fitting.Di Li Cray process mixture model is one of model most popular in Bayes's nonparametric model, and it is potential class models, is commonly used in clustering problem.

Summary of the invention

The invention provides a kind of TAC clustering method based on Di Li Cray process mixture model, the problems such as the poor fitting easily occurred when can solve Model Selection difficulty and existing clustering method modeling and over-fitting.

Based on a TAC clustering method for Di Li Cray process mixture model, comprise the steps:

(1) initializing set class number K, focuses parameters α, class degree of separation correlation parameter ss, s0, and the value of the degree of freedom v of the special covariance priori in inverse Vichy;

(2) based on the initialization information in step (1), initialization is carried out to the classification that every bar TAC belongs to, set up Di Li Cray process mixture model and calculate the relevant information q determining each classification in mixture model;

(3) for all TAC, successively every bar TAC shifted out from mixture model and calculate its conditional probability belonging to each classification and new classification;

(4) according to described conditional probability, this TAC is sampled selection classification again, then by its add-back mixture model and Renewal model parameter;

(5) travel through all TAC according to step (3) and (4) and take that iteration once as, iterating until class number K no longer changes, and preserve final result.

Carry out initialization to the classification that every bar TAC belongs in described step (2), calculate the relevant information q determining each classification in mixture model, detailed process is as follows:

2.1 represent the value of the class belonging to all TAC with z, z _ifor the element of i-th in z, z _ivalue obtained by generation random natural number, its span is 1≤z _i≤ K, i are natural number and 1≤i≤N, N is the sum of TAC;

Relevant information q described in 2.2 comprises three components: qn, qo and qm; For arbitrary classification k, its relevant information q (k) of initialization, that is:

q·n(k)＝q·o(k)＝q·m(k)＝0

Wherein: q (k) represents the relevant information of kth class, qn (k), qo (k) and qm (k) correspond to three components in relevant information q (k), and k is class number and is natural number, 1≤k≤K;

2.3 for arbitrary TAC, carries out renewal rewards theory by following relational expression to the relevant information q of each classification:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q·n(z _i)′＝q·n(z _i)+1

q·o(z _i)′＝q·o(z _i)+x(i)

q·m(z _i)′＝q·m(z _i)+mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm be all elements in m and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, qn (z _i), qo (z _i) and qm (z _i) correspond to the front z of renewal _irelevant information q (the z of class _i) in three components, qn (z _i) ', qo (z _i) ' and qm (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC;

2.4 travel through every bar TAC according to step 2.3, and the result after traversal upgrades is finally as the relevant information qq of each classification in mixture model.

Calculate the conditional probability that every bar TAC belongs to each classification and new classification in described step (3), detailed process is as follows:

First current TAC shifts out by 3.1 from mixture model, and upgrades the relevant information q of each classification in mixture model by following relational expression:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q·n(z _i)′＝q·n(z _i)+1

q·o(z _i)′＝q·o(z _i)+x(i)

q·m(z _i)′＝q·m(z _i)+mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm and m be middle all elements and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, qn (z _i), qo (z _i) and qm (z _i) correspond to the front z of renewal _iclass relevant information q (z _i) in three components, qn (z _i) ', qo (z _i) ' and qm (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC;

After relational expression in 3.2 completing steps 3.1 calculates, if qn is (z _i) ' value be 0, then current z is described _ithe class corresponding to value in element number be 0, namely in such not containing any element, at this moment need the class of deleting this sky, if current z _ivalue be k, need upgrade K, qn, qo, qm and z _ivalue, concrete update mode is:

K′＝K-1；

The qn of qn '=delete a kth element

The qo of qo '=delete a kth element

The qm of qm '=delete a kth element

Z ' _i=z _i-1 (if z _i> k)

Wherein: K, qn, qo, qm, z _iinformation before corresponding renewal, K ', qn ', qo ', qm ', z ' _iinformation after corresponding renewal, for z _irenewal need to travel through each TAC, namely i is natural number and its value is the sum of TAC from 1 to N, N; If after the relational expression in completing steps 3.1 calculates, if qn is (z _i) ' value be not 0, then without the need to this single stepping;

Then 3.3 calculate this TAC by following relational expression, and to belong to the condition of each classification and new classification general

p1＝log([q·n，α])

$p2\left(k\right)=p1\left(k\right)=(q.n(k)+1)*\frac{d}{2}*\log \left(\mathrm{\π}\right)-\frac{d}{2}*\log \left(\frac{{\mathrm{ss}}^2}{s{0}^2}+1\right)-$

$(v+1)*\mathrm{\Σ}\left(\log \left(diag\left(chol2\left(chol1\left(chol\left({\mathrm{ss}}^2*I_d*v\right),x\left(i\right)\right),\frac{x\left(i\right)}{\sqrt{\frac{{\mathrm{ss}}^2}{s{0}^2}+1}}\right)\right)\right)\right)$

$+q.n\left(k\right)*\frac{d}{2}*log\left(\mathrm{\π}\right)+\frac{d}{2}*log\left(\frac{{\mathrm{ss}}^2}{s{0}^2}\right)$

$+v*\mathrm{\Σ}\left(\log \left(diag\left(chol2\left(chol\left({\mathrm{ss}}^2*I_d*v\right),O_{d,1}\right)\right)\right)\right)$

p3＝exp(p2-max(p2))

p＝p3/sum(p3)

Wherein: x (i) is i-th TAC and current TAC, d represents the dimension of TAC, p is the vector that i-th TAC belongs to the conditional probability composition of each class, namely element p (k) wherein represents that i-th TAC belongs to the conditional probability of kth class, wherein, last element representation i-th TAC in p belongs to the conditional probability of a new class, k is class number and is natural number, 1≤k≤k+1, α is focuses parameters, s0 and ss is class degree of separation correlation parameter, v is the degree of freedom of the special covariance priori in inverse Vichy, qn (k) is one of them component in kth class relevant information q (k), O _{d, 1}be line number to be d columns be 1 null vector, chol (A) is Cholesky analytic function, if matrix A is symmetric positive definite, then Cholesky decomposes product matrix A being resolved into a lower triangular matrix and upper triangular matrix, if represent upper triangular matrix with R, then lower triangular matrix is its transposition, i.e. A=R ' R, if R=chol (A) is the Cholesky factor of A, so chol1 (R, X) result that function obtains is the Cholesky factor of (A-X*X '), chol2 (R, X) result that function obtains is the Cholesky factor of (A+X ' * X), chol1 (R, and chol2 (R X), X) function all only uses the diagonal angle of R and upper triangle element, the lower triangle element of R is all left in the basket, the effect of diag (X) function is that the element on the diagonal line of X is taken out the new vector of composition one, here summation symbol ∑ is sued for peace to each element in vector in bracket.

Again to sample selection classification to this TAC according to described conditional probability in described step (4), then by its add-back mixture model and Renewal model parameter, detailed process is as follows:

Each element in the p obtained in 4.1 pairs of steps 3.3 is sampled, and obtains sampling results, and the conditional probability drawing a kth element in p is p (k);

If 4.2 what draw is last element, i.e. (K+1) individual element in p, then need the class that structure one is new, concrete operations mode is as follows:

K′＝K+1；

q·n′＝[q·n，0]

q·o′＝[q·o，0]

q·m′＝[q·m，0]

z′ _i＝K′

Wherein, the information before K, qn, qo, qm correspondence upgrades, the information after K ', qn ', qo ', qm ' correspondence upgrade; If what draw is not last element in p, then without the need to this single stepping;

4.3 then by current TAC add-back mixture model, and at this moment need to revise three component qn of q, the value of qo, qm again, concrete modification mode is as follows:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q·n(z _i)′＝q·n(z _i)+1

q·o(z _i)′＝q·o(z _i)+x(i)

q·m(z _i)′＝q·m(z _i)+mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm and m be middle all elements and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, qn (z _i), qo (z _i) and qm (z _i) correspond to the front z of renewal _irelevant information q (the z of class _i) in three components, qn (z _i) ', qo (z _i) ' and qm (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC.

The span of described initialization class number K is 100 ~ 100000, the span of focuses parameters α is 0.001 ~ 10000, the span of class degree of separation correlation parameter ss and s0 is 0.01 ~ 1000, and the span of the degree of freedom v of the special covariance priori in inverse Vichy is 10 ~ 10000.

The present invention carries out cluster by using Di Li Cray process mixture model to TAC, effectively solve the problem of carrying out cluster when the unknown of class number, and the complexity of Di Li Cray mixture model can increase along with the increase of the data bulk obtained, this is the advantage not available for other clustering algorithms.

Accompanying drawing explanation

Fig. 1 is the steps flow chart schematic diagram of TAC clustering method of the present invention.

Fig. 2 be true value corresponding TAC figure; Wherein horizontal ordinate is frame number, and ordinate is the grey scale change of each pixel with frame number.

The cluster result schematic diagram that Fig. 3 (a) is true value; Wherein the pixel of the position of different gray scale represents different classes, and namely the TAC representated by pixel of same gray scale belongs to same class.

Fig. 3 (b) is for class number in cluster process is with the change schematic diagram of iterations.

Embodiment

In order to more specifically describe the present invention, below in conjunction with the drawings and the specific embodiments, TAC clustering method of the present invention is described in detail.

As shown in Figure 1, the present invention is based on the TAC clustering method of Di Li Cray process mixture model, comprise the steps:

S1. initialization parameters: need initialized parameter to have class number K, focuses parameters α, class degree of separation correlation parameter ss, s0, and the value of the degree of freedom v of the special covariance priori in inverse Vichy;

S2. initialization DP mixture model: initialization is carried out to the classification that every bar TAC belongs to, calculate the relevant information q determining each classification in mixture model, detailed process is as follows:

2.1 represent the value of the class belonging to all TAC with z, z _ifor the element of i-th in z, z _ivalue obtained by generation random natural number, its span is 1≤z _i≤ K, i are natural number and 1≤i≤N, N is the sum of TAC;

Relevant information q described in 2.2 comprises three components: qn, qo and qm; For arbitrary classification k, its relevant information q (k) of initialization, namely

q·n(k)＝q·o(k)＝q·m(k)＝0

Wherein: q (k) represents the relevant information of kth class, qn (k), qo (k) and qm (k) correspond to three components in relevant information q (k), and k is class number and is natural number, 1≤k≤K;

2.3 for arbitrary TAC, carries out renewal rewards theory by following relational expression to the relevant information q of each classification:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q·n(z _i)′＝q·n(z _i)+1

q·o(z _i)′＝q·o(z _i)+x(i)

q·m(z _i)′＝q·m(z _i)+mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm be all elements in m and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, qn (z _i), qo (z _i) and qm (z _i) correspond to the front z of renewal _irelevant information q (the z of class _i) in three components, qn (z _i) ', qo (z _i) ' and qm (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC;

2.4 travel through every bar TAC according to step 2.3, and the result after traversal upgrades is finally as the relevant information qq of each classification in mixture model.

S3. design conditions probability: calculate the conditional probability that every bar TAC belongs to each classification and new classification, detailed process is as follows:

First current TAC shifts out by 3.1 from mixture model, and upgrades the relevant information q of each classification in mixture model by following relational expression:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q·n(z _i)′＝q·n(z _i)+1

q·o(z _i)′＝q·o(z _i)+x(i)

q·m(z _i)′＝q·m(z _i)+mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm and m be middle all elements and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, qn (z _i), qo (z _i) and qm (z _i) correspond to the front z of renewal _iclass relevant information q (z _i) in three components, qn (z _i) ', qo (z _i) ' and qm (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC;

After 3.2 relational expressions completed in 3.1 calculate, if qn is (z _i) ' value be 0, then current z is described _ithe class corresponding to value in element number be 0, namely in such not containing any element, at this moment need the class of deleting this sky, if current z _ivalue be k, need upgrade K, qn, qo, qm and z _ivalue, concrete update mode is:

K′＝K-1；

The qn of qn '=delete a kth element

The qo of qo '=delete a kth element

The qm of qm '=delete a kth element

Z ' _i=z _i-1 (if z _i> k)

Wherein: K, qn, qo, qm, z _iinformation before corresponding renewal, K ', qn ', qo ', qm ', z ' _iinformation after corresponding renewal, for z _irenewal need to travel through each TAC, namely i is natural number and its value is the sum of TAC from 1 to N, N; If after the relational expression completed in 3.1 calculates, if qn is (z _i) ' value be not 0, then without the need to this single stepping;

Then 3.3 calculate by following relational expression the conditional probability that this TAC belongs to each classification and new classification:

p1＝log([q·n，α])

$p2\left(k\right)=p1\left(k\right)=(q.n(k)+1)*\frac{d}{2}*\log \left(\mathrm{\π}\right)-\frac{d}{2}*\log \left(\frac{{\mathrm{ss}}^2}{s{0}^2}+1\right)-$

$(v+1)*\mathrm{\Σ}\left(\log \left(diag\left(chol2\left(chol1\left(chol\left({\mathrm{ss}}^2*I_d*v\right),x\left(i\right)\right),\frac{x\left(i\right)}{\sqrt{\frac{{\mathrm{ss}}^2}{s{0}^2}+1}}\right)\right)\right)\right)$

$+q.n\left(k\right)*\frac{d}{2}*log\left(\mathrm{\π}\right)+\frac{d}{2}*log\left(\frac{{\mathrm{ss}}^2}{s{0}^2}\right)$

$+v*\mathrm{\Σ}\left(\log \left(diag\left(chol2\left(chol\left({\mathrm{ss}}^2*I_d*v\right),O_{d,1}\right)\right)\right)\right)$

p3＝exp(p2-max(p2))

p＝p3/sum(p3)

Wherein: x (i) is i-th TAC and current TAC, d represents the dimension of TAC, p is the vector that i-th TAC belongs to the conditional probability composition of each class, namely element p (k) wherein represents that i-th TAC belongs to the conditional probability of kth class, wherein, last element representation i-th TAC in p belongs to the conditional probability of a new class, k is class number and is natural number, 1≤k≤K+1, α is focuses parameters, s0 and ss is class degree of separation correlation parameter, v is the degree of freedom of the special covariance priori in inverse Vichy, qn (k) is one of them component in kth class relevant information q (k), O _{d, 1}be line number to be d columns be 1 null vector, chol (A) is Cholesky analytic function, if matrix A is symmetric positive definite, then Cholesky decomposes product matrix A being resolved into a lower triangular matrix and upper triangular matrix, if represent upper triangular matrix with R, then lower triangular matrix is its transposition, i.e. A=R ' R, if R=chol (A) is the Cholesky factor of A, so chol1 (R, X) result that function obtains is the Cholesky factor of (A-X*X '), chol2 (R, X) result that function obtains is the Cholesky factor of (A+X ' * X), chol1 (R, and chol2 (R X), X) function all only uses the diagonal angle of R and upper triangle element, the lower triangle element of R is all left in the basket, the effect of diag (X) function is that the element on the diagonal line of X is taken out the new vector of composition one, here summation symbol ∑ is sued for peace to each element in vector in bracket.

S4. sampling Renewal model: this TAC is sampled selection classification again according to the conditional probability calculated in S3, then by its add-back mixture model and Renewal model parameter, detailed process is as follows:

Each element in the p obtained in 4.1 couples 3.3 is sampled, and obtains sampling results, and the conditional probability drawing a kth element in p is p (k).

If 4.2 what draw is last element, i.e. (K+1) individual element in p, then need the class that structure one is new, concrete operations mode is as follows:

K′＝K+1；

q·n′＝[q·n，0]

q·o′＝[q·o，0]

q·m′＝[q·m，0]

z′ _i＝K′

Wherein, the information before K, qn, qo, qm correspondence upgrades, the information after K ', qn ', qo ', qm ' correspondence upgrade; If what draw is not last element in p, then without the need to this single stepping;

4.3 then by current TAC add-back mixture model, and at this moment need to revise three component qn of q, the value of qo, qm again, concrete modification mode is as follows:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q·n(z _i)′＝q·n(z _i)+1

q·o(z _i)′＝q·o(z _i)+x(i)

q·m(z _i)′＝q·m(z _i)+mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm and m be middle all elements and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, qn (z _i), qo (z _i) and qm (z _i) correspond to the front z of renewal _irelevant information q (the z of class _i) in three components, qn (z _i) ', qo (z _i) ' and qm (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC.

S5. repeat step S3 and S4, until the class number K of data no longer changes, now iteration stopping, preserves the value of q, K, z.

We verify practicality and the reliability of present embodiment by experiment true value (groundtruth) being carried out to cluster below.True value is made up of 18 two field pictures, and the size of every two field picture is 128 pixel * 128 pixels, and the gray-scale value of the same pixel on every two field picture can form ten octuple vectors, i.e. TAC, we can draw TAC curve, as shown in Figure 2, horizontal ordinate is frame number, and ordinate is gray-scale value.As shown in Figure 3, the pixel of different gray areas belongs to different classes to the operation result of program, and namely the pixel of same grayscale belongs to same class, the corresponding TAC of each pixel.The program of present embodiment can obtain correct cluster result with the probability (because have the composition of random sampling inside in program, therefore not can obtain correct cluster result, this is normal phenomenon) of more than 95% at every turn.

Claims (5)

1., based on a TAC clustering method for Di Li Cray process mixture model, comprise the steps:

(1) value of the degree of freedom v of initializing set class number K, focuses parameters α, the special covariance priori of class degree of separation correlation parameter ss and s0 and inverse Vichy;

(2) based on the initialization information in step (1), initialization is carried out to the classification that every bar TAC belongs to, set up Di Li Cray process mixture model and calculate the relevant information q determining each classification in mixture model;

(3) for all TAC, successively every bar TAC shifted out from mixture model and calculate its conditional probability belonging to each classification and new classification;

(4) according to described conditional probability, this TAC is sampled selection classification again, then by its add-back mixture model and Renewal model parameter;

(5) travel through all TAC according to step (3) and (4) and take that iteration once as, iterating until class number K no longer changes, and preserve final result.

2. the TAC clustering method based on Di Li Cray process mixture model according to claim 1, it is characterized in that: in described step (2), initialization is carried out to the classification that every bar TAC belongs to, calculate the relevant information q determining each classification in mixture model, detailed process is as follows:

2.1 represent the value of the class belonging to all TAC with z, z _ifor the element of i-th in z, z _ivalue obtained by generation random natural number, its span is 1≤z _i≤ K, i are natural number and 1≤i≤N, N is the sum of TAC;

Relevant information q described in 2.2 comprises three components: q.n, q.o and q.m; For arbitrary classification k, its relevant information q (k) of initialization, that is:

q.n(k)＝q.o(k)＝q.m(k)＝0

Wherein: q (k) represents the relevant information of kth class, q.n (k), q.o (k) and q.m (k) correspond to three components in relevant information q (k), and k is class number and is natural number, 1≤k≤K;

2.3 for arbitrary TAC, carries out renewal rewards theory by following relational expression to the relevant information q of each classification:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q.n(z _i)′＝q.n(z _i)+1

q.o(z _i)′＝q.o(z _i)+x(i)

q.m(z _i)′＝q.m(z _i)+mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm be all elements in m and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, q.n (z _i), q.o (z _i) and q.m (z _i) correspond to the front z of renewal _irelevant information q (the z of class _i) in three components, q.n (z _i) ', q.o (z _i) ' and q.m (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC;

2.4 travel through every bar TAC according to step 2.3, and the result after traversal upgrades is finally as the relevant information qq of each classification in mixture model.

3. the TAC clustering method based on Di Li Cray process mixture model according to claim 1, is characterized in that: calculate the conditional probability that every bar TAC belongs to each classification and new classification in described step (3), detailed process is as follows:

First current TAC shifts out by 3.1 from mixture model, and upgrades the relevant information q of each classification in mixture model by following relational expression:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q.n(z _i)′＝q.n(z _i)-1

q.o(z _i)′＝q.o(z _i)-z(i)

q.m(z _i)′＝q.m(z _i)-mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm and m be middle all elements and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, q.n (z _i), q.o (z _i) and q.m (z _i) correspond to the front z of renewal _iclass relevant information q (z _i) in three components, q.n (z _i) ', q.o (z _i) ' and q.m (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC;

After calculating in 3.2 completing steps 3.1, if q.n is (z _i) ' value be 0, then current z is described _ithe class corresponding to value in element number be 0, namely in such not containing any element, at this moment need the class of deleting this sky, if current z _ivalue be k, need upgrade K, q.n, q.o, q.m and z _ivalue, concrete update mode is:

K′＝K-1；

The q.n of q.n '=delete a kth element

The q.o of q.o '=delete a kth element

The q.m of q.m '=delete a kth element

Z ' _i=z _i-1 (if z _i> k)

Wherein: K, q.n, q.o, q.m, z _iinformation before corresponding renewal, K ', q.n ', q.o ', q.m ', z ' _iinformation after corresponding renewal, for z _irenewal need to travel through each TAC, namely i is natural number and its value is the sum of TAC from 1 to N, N;

If after the calculating in completing steps 3.1, q.n (z _i) ' value be not 0, then without the need to carrying out aforesaid operations;

Then 3.3 calculate by following relational expression the conditional probability that this TAC belongs to each classification and new classification:

p1＝log([q.n，α])

$\begin{array}{c}p2\left(k\right)=p1\left(k\right)-(q.n(k)+1)*\frac{d}{2}*\log \left(\mathrm{\π}\right)-\frac{d}{2}*\log (\frac{{\mathrm{ss}}^2}{s{0}^2}+1)-\\ (v+1)*\mathrm{\Σ}\left(\log \right(diag\left(chol2\right(chol1\left(chol\right({\mathrm{ss}}^2*I_d*v),x(i\left)\right),\frac{x\left(i\right)}{\sqrt{\frac{{\mathrm{ss}}^2}{s{0}^2}+1}}\left)\right)\left)\right)\\ +q.n\left(k\right)*\frac{d}{2}*\log \left(\mathrm{\π}\right)+\frac{d}{2}*\log \left(\frac{{\mathrm{ss}}^2}{s{0}^2}\right)\end{array}$

$+v*\mathrm{\Σ}\left(log\right(diag\left(chol2\left(chol\left({\mathrm{ss}}^2*I_d*v\right),O_{d,1}\right)\right)\left)\right)$

p3＝exp(p2-max(p2))

p＝p3/sum(p3)

Wherein: x (i) is i-th TAC and current TAC, d represents the dimension of TAC, p is the vector that i-th TAC belongs to the conditional probability composition of each class, namely element p (k) wherein represents that i-th TAC belongs to the conditional probability of kth class, wherein, last element representation i-th TAC in p belongs to the conditional probability of a new class, k is class number and is natural number, 1≤k≤K+1, α is focuses parameters, s0 and ss is class degree of separation correlation parameter, v is the degree of freedom of the special covariance priori in inverse Vichy, q.n (k) is one of them component in kth class relevant information q (k), O _{d, 1}be line number to be d columns be 1 null vector, chol (A) is Cholesky analytic function, if matrix A is symmetric positive definite, then Cholesky decomposes product matrix A being resolved into a lower triangular matrix and upper triangular matrix, if represent upper triangular matrix with R, then lower triangular matrix is its transposition, i.e. A=R ' R, if R=chol (A) is the Cholesky factor of A, so chol1 (R, X) result that function obtains is the Cholesky factor of (A-X*X '), chol2 (R, X) result that function obtains is the Cholesky factor of (A+X ' * X), chol1 (R, and chol2 (R X), X) function all only uses the diagonal angle of R and upper triangle element, the lower triangle element of R is all left in the basket, the effect of diag (X) function is that the element on the diagonal line of X is taken out the new vector of composition one, summation symbol ∑ in above formula is sued for peace to each element in vector in bracket.

4. the TAC clustering method based on Di Li Cray mixture model according to claim 3, it is characterized in that: according to described conditional probability, this TAC is sampled selection classification again in described step (4), then by its add-back mixture model and Renewal model parameter, detailed process is as follows:

Each element in the p obtained in 4.1 pairs of steps 3.3 is sampled, and obtains sampling results, and the conditional probability drawing a kth element in p is p (k);

If 4.2 what draw is last element, i.e. (K+1) individual element in p, then need the class that structure one is new, concrete operations mode is as follows:

K′＝K+1；

q.n′＝[q.n，0]

q.o′＝[q.o，0]

q.m′＝[q.m，0]

z′ _i＝K′

Wherein, the information before K, q.n, q.o, q.m correspondence upgrades, the information after K ', q.n ', q.o ', q.m ' correspondence upgrade;

If what draw is not last element in p, then without the need to aforesaid operations;

4.3 then by current TAC add-back mixture model, and at this moment need to revise three component q.n of q, the value of q.o, q.m again, concrete modification mode is as follows:

$mm={\mathrm{\Σ}}_{k=1}^{M}m\left(k\right)$

q.n(z _i)′＝q.n(z _i)+1

q.o(z _i)′＝q.o(z _i)+x(i)

q.m(z _i)′＝q.m(z _i)+mm

Wherein: x (i) is i-th TAC and current TAC, m is the vector of all elements composition not being 0 in i-th TAC, m (k) is the element of the kth in m, M is the dimension of this vector, namely be not the number of all elements of 0 in current TAC, mm and m be middle all elements and, z _irepresent the value of the class belonging to i-th TAC, its span is 1≤z _i≤ K, q.n (z _i), q.o (z _i) and q.m (z _i) correspond to the front z of renewal _irelevant information q (the z of class _i) in three components, q.n (z _i) ', q.o (z _i) ' and q.m (z _iz after) ' correspond to upgrades _iclass relevant information q (z _i) in three components, i is natural number and 1≤i≤N, N is the sum of TAC.

5. the TAC clustering method based on Di Li Cray process mixture model according to claim 1, it is characterized in that: in described step (1), the span of initialization class number K is 100 ~ 100000, the span of focuses parameters α is 0.001 ~ 10000, the span of class degree of separation correlation parameter ss and s0 is 0.01 ~ 1000, and the span of the degree of freedom v of the special covariance priori in inverse Vichy is 10 ~ 10000.